citation-trace

citation-trace

Citation graph exploration for a seed paper.

Invariant: always ask, never default

Any parameter below the user did not specify MUST be asked before doing work. Bundle related questions.

direction has no default and must always be confirmed. The user flagged this specifically — even if the seed paper and every other parameter is clear, STOP and ask for direction.

Required parameters — ASK if missing

1. Seed paper — arxiv ID, DOI, Semantic Scholar ID, or Zotero citation key.
2. Direction — forward (papers citing the seed), backward (papers the seed cites), or both. Never assume.
3. Depth — hops from the seed (1 = direct; 2 = also citations-of-citations).
4. Cap per direction — a max results limit (e.g., 20 / 50 / 100).

To confirm — ASK, do not assume

• Relevance filter — keywords, topics, year range, or none?
• Cross-reference against Zotero — flag results already in the user's library? If yes, use mcp__zotero__zotero_search_items / similar to check.
• Cross-reference against vault — flag results that match existing notes in 3-Notes/? If yes, load /vaultdb.
• Grouping of results — by relevance / by year / by venue / by topical cluster?
• Save output as a note? Default is NO (inline report). If yes, hand off to /paper-notes.

Data sources

• Semantic Scholar is the primary citation-graph source. Endpoints:
  • Backward: https://api.semanticscholar.org/graph/v1/paper/<ID>/references
  • Forward: https://api.semanticscholar.org/graph/v1/paper/<ID>/citations
  • Useful fields: ?fields=title,authors,year,venue,abstract,citationCount,externalIds
• arxiv — metadata for arxiv-hosted results (WebFetch arxiv.org/abs/<id>).
• Zotero MCP — cross-ref against the user's library.

Workflow

1. Confirm every parameter (above). Do not skip.
2. Fetch seed metadata and echo it back: "Seed: , <first author et al.>, <year>. Proceed?"
3. Fetch citations in confirmed direction(s) up to confirmed depth and cap.
4. Apply relevance filter per confirmation.
5. Cross-reference with Zotero / vault per confirmation.
6. Group per confirmed strategy.
7. Report.

Output format

## Seed
<Title> — <Authors>, <Year>, <Venue>

## Parameters (as confirmed)
direction=<...>, depth=<N>, cap=<M>, filter=<...>, grouping=<...>

## Results by <confirmed grouping>

### <Group 1> (<count>)

- **<Title>** — <Authors>, <Year>, <Venue>. Cited <N>. [in library: yes/no] [in vault: yes/no]
  Why relevant: <1-line rationale tied to the seed's content>

...

## Suggestions
- Strongest prior-art seeds: <top 3>
- Strongest extension seeds: <top 3>
- Consider `/paper-triage` on: <top 2-3>
- Consider `/lit-review` if any cluster crosses <threshold>

Guardrails

• Never fabricate citation counts. If an API omits the field, omit it in output.
• If Semantic Scholar rate-limits or the seed isn't found, report honestly — do not invent a graph.
• Do not persist results to the vault without explicit confirmation.

$ARGUMENTS